Why Running AI Agents in Production Requires a New Infrastructure Playbook

Platform engineers often assume AI agents can be deployed like standard containerized microservices, but this approach breaks down quickly in production environments. Unlike typical services, agent requests can run for minutes, carry state across multiple tool calls, and return HTTP 200 status codes even when their outputs are semantically wrong. Common Kubernetes defaults — such as liveness probes with 30-second thresholds and CPU-based autoscalers — are poorly suited to agents that may legitimately pause for 40–90 seconds during a reasoning step. Proper agent infrastructure requires separating process-liveness checks from forward-progress monitoring, setting timeout budgets per tool-call chain rather than per request, and treating session isolation as an architectural concern. The core argument is that agent infrastructure demands its own discipline, distinct from the microservice patterns that platform teams have relied on for the past decade.